Deep learning classification of longitudinally defined 24-2 central- and peripheral-onset visual field phenotypes in glaucoma using baseline fundus photographs

Abstract We aimed to determine whether a deep learning (DL) model trained on baseline color fundus photographs can identify the dominant locus of visual field (VF) damage in glaucoma, as defined by longitudinal VF follow-up. We performed retrospective analysis of a longitudinal cohort study on 1,464 patients with early glaucoma (baseline mean deviation > − 6 dB) followed for at least 5 years with a minimum of five reliable VF examinations. Eyes were classified into operationally defined central-onset or peripheral-onset VF damage patterns based on longitudinally consistent VF patterns. A retinal foundation model RETFound was fine-tuned in two configurations, unimodal (fundus image only) and multimodal (fundus image + gender via late fusion), to classify central-onset versus peripheral-onset phenotype. Model performance was assessed on a held-out test set. A total of 1,088 eyes from 871 patients were finally included in the analysis; 168 (15.4%) were classified as central-onset and 920 (84.6%) as peripheral-onset phenotype. At baseline, 175 eyes (16.1%) were preperimetric, whereas 913 eyes (83.9%) had a detectable VF defect. The mean follow-up duration was 9.0 ± 2.7 years and did not differ between groups. The multimodal RETFound model achieved the highest AUROC of 0.867, followed by the unimodal model (0.840) and ConvNeXt (0.839), while other backbone architectures ranged from 0.641 to 0.721. In conclusion, DL applied to baseline fundus photographs identified longitudinally defined patterns of central versus peripheral VF vulnerability in glaucoma.